LuxTrace - Indoor Positioning Using Building Illumination · perposition for energy eﬃciency and...

LuxTrace - Indoor Positioning Using BuildingIllumination

Julian Randall1, Oliver Amft1, Jurgen Bohn2, and Martin Burri1

1 Wearable Computing Lab., ETH Zurich, Switzerland, [email protected],[email protected],[email protected],

2 Institute for Pervasive Computing, ETH Zurich, Switzerland, [email protected]

Summary. Tracking location is challenging due to the numerous constraints ofpractical systems including, but not limited to, global cost, device volume andweight, scalability and accuracy; these constraints are typically more severe for sys-tems that should be wearable and used indoors.

We investigate the use of wearable solar cells to track changing light conditions(a concept that we named LuxTrace) as a source of user displacement and activitydata. We evaluate constraints of this approach and present results from an experi-mental validation of displacement and activity estimation. The results indicate thata distance estimation accuracy of 21 cm (80% quantile) can be achieved.

A simple method to combine LuxTrace with complementary absolute locationestimation methods is also presented. We apply carpet-like distributed RFID tagsto demonstrate online learning of new lighting environments.

1 Introduction

The LuxTrace project investigates the opportunities for using solar cells tosimultaneously collect energy and track light level. As light conditions arelocation dependent, this radiance data can be used as context information tosupport the position tracking of a user indoors.

Advantages of extracting context data from indoor lights are that thisinfrastructure may be used without modification (i.e. the technology is min-imally invasive) and this source of context data is generally untapped. Solarcells as sensors for wearable computing are attractive firstly as their low num-ber of “pixels” (e.g. 1) implies that limited processing power may be sufficient.Secondly, as solar cells can be produced in thin (< 1 mm) low weight layers ona flexible substrate in a number of colours they better match garment speci-fications than other sensing devices. A third benefit of solar cells is that theycan be used as a sensor and an energy harvesting device, thus contributing tothe restricted energy available on the body.

2 Randall, Amft, Bohn and Burri

The limitations of existing technologies are related to a number of factorsincluding coverage, cost and location estimation accuracy. Satellite navigationfor indoor application is unsatisfactory in all three categories given that it isgenerally not available indoors, is relatively expensive and the error of GlobalPositioning System (GPS) for example is of the order of 15-20m [1]. This er-ror can be reduced to 1-2 m by error correction using European GeostationaryNavigation Overlay Service [2]. The European Galileo [3] system, planned tobe commercially available by 2008, targets a global accuracy on the horizontalplane of up to 4m for safety of life services and mass market applications, andup to 1 m for commercial applications. Galileo may offer horizontal accuracyin the centimetre-range by means of locally augmented signals [4]. Numer-ous other location tracking technologies exist that provide absolute positionestimations, e.g. radio frequency identification (RFID) tags [5] or wirelesscommunication systems [6, 7]. Generally these technologies rely on support-ing infrastructure. Relative position technologies include ultrasound [8] andinertial sensors, e.g. [9]. These systems suffer from continuous drift since nore-calibration mechanism is integrated. Some of these technologies have beendeveloped in badge form, but generally have not been fully integrated intoclothing.

Our investigations evaluate the potential for using solar modules aloneas a low-cost alternative for estimating displacement and collecting contextinformation indoors by measuring the radiant intensity (e.g. from overheadlighting). We expect that solar cells can be integrated into garments, e.g. onthe user’s shoulder. We present a theoretical model for ideal light sourceswithout reflector and diffuser. We validate this model with experiments in anoffice corridor with evenly spaced fluorescent tubes using a trolley equippedwith solar cells. Using a trained model, we estimate displacement with anerror of 21 cm (confidence 80%) on straight trajectories using the data froma single solar cell. In order to determine the impact of user movement on thepositioning accuracy, we then report experimental results of a user on a tread-mill under fixed lighting. We show that light intensity varies significantly morewith the horizontal displacement related to walking than the cyclic motion ofthe shoulder.

Single technologies have the strategic disadvantages that if the sensor sys-tem fails, is locally unavailable or suffers from drift (e.g. due to dead reckoning)the positioning service may fail or be unreliable. This supports the case forusing multiple technologies. Since LuxTrace is a relative positioning approachbased on tracking of individual lights the combination with an absolute po-sitioning reference can be advantageous. We present a first concept to learnnew lighting environments online by using an absolute position informationwhen it is available. Applications may include corridor intersections wherelighting conditions change or areas with mixed daylight and electrical light,e.g. in large rooms with multiple light sources.

Other investigations that have combined two technologies include the workof Lee et al. [10] and Randell and Muller [11]. However, these combinations

LuxTrace - Indoor Positioning Using Building Illumination 3

are non-redundant i.e. these systems require data from both sensors simul-taneously. The HeyWow project [12] presented the idea of combining sensingtechnologies in parallel i.e. with redundancy. Similar approaches with sensorfusion technology to ensure availability and to increase accuracy have beenmade in [13]. Angerman et al. [14] suggests that sensor measurements canbe considered as a probability distribution function. Generalisations of sensorfusion technology have been presented in the COMPASS project [15] and inthe Location Stack project [16].

We have selected RFID tags distributed in carpet-like foils on the flooras complementary technology since this technology offers an absolute refer-ence positioning service with low-maintenance and low cost [17]. Moreoverthe memory capacity of the RFID tags could be used for providing additionallocation-dependent information in-situ. We conducted initial experiments tostudy the combined system with a similar trolley and corridor to maintaincomparability with the original experiments. The combined system collectslocation-dependent information from the RFID tags. Apart from online pro-cessing of position estimates, the combined system dynamically adapts tochanges in the installed overhead lighting by means of an online model learn-ing procedure. The results of the combined system indicate that a similarperformance can be maintained in a more practical setting. Hence, the modellearning approach is a feasible strategy for displacement estimation.

The paper follows the structure of our investigations. Firstly, we presentour sensing approach and discuss issues related to it (Section 2). In Section 3a general light emitter model for indoor light tracking is developed and evalu-ated. Section 4 presents our displacement estimation experiments and results.In Section 5 we present the initial experiments and results for the combinedsystem. These results are assessed in the Discussion (Section 6). Finally weindicate further research directions in Section 7.

2 Solar Cell Navigation Approach

In our sensing approach we use a solar module alone as a low-cost meansto estimate displacement and collect context information indoors. We expectthat solar cells can detect changes in the radiant intensity (e.g. from over-head lighting) when moved. Moreover this sensor can be easily integrated intogarments, e.g. at the user’s shoulder. Such modules can have similar char-acteristics to textiles e.g low cost (∼2 US$), low weight (∼20 g), low volume(∼2 cm3) and with a range of colours. A potential concept is shown in Figure 1in which the flexible solar module system on the shoulder (1), transmits oneor more RF pulses only when there is sufficient energy to do so i.e. beneath alight source (2). This data is collected and processed on-body, e.g. by a beltworn computer (3) [18].

The environment is assumed to be static as the lights in the corridor arealways on during office hours [19]. Whilst it is necessary to process the data


Fig. 1. LuxTrace concept: a wearable solar cells (1) attached to the user’s shoulder,transmits measurements, e.g. via RF pulses (2) and on-body data processing, e.g.in a belt (3).

from the solar modules, their relatively low bit rate is well adapted to on-bodyprocessing such as with a body worn computer.

We investigate how light data can be used to support indoor navigationsystems. The issues relevant to LuxTrace can be categorised into three mainareas: 1) the light sources providing the building illumination, that are usedfor tracking, 2) the environmental influences on the tracking of light sourcesand 3) the effects related to the user wearing the system. To demonstrate thescalability of our approach, we discuss in this section important items relatedto each area.

2.1 Light Source Related Aspects

Distance and light angle: In an ideal case radiant intensity received at thesolar cell declines at 1/r2 with increasing distance r between the solar celland light source(s). Intensity is also reduced if the plane of the solar cell isnot perpendicular to the incident light. This effect can support the lightsource tracking approach, since a maximisation of the voltage differencebetween “peaks” (maximum voltage) when the receiver is perpendicularunder a light source and “valleys” (minimum voltage) when the receiveris between light sources is intended. The experimental conditions used inthis work are outlined in Section 4.1 and a theoretical assessment is givenin Section 3.

Light installation, diffusers and superposition: The way light sources are in-stalled can affect measured radiant intensity. This can be related to thereflective and spectral properties of the objects around the light sourcesuch as diffusers and reflectors. Another aspect of light installation is en-countered when the distance between two light sources is reduced: by


superposition a light intensity maximum between two light sources can becreated. Algorithms may misinterpret this intensity peak as a light source.Fortunately normal building lighting schemes avoid such pronounced su-perposition for energy efficiency and visual comfort reasons. Exceptions in-clude areas where higher light level is required, e.g. work benches. Our ex-periments have been made with a standard diffuser type (see Section 4.1).Pronounced superposition did not occur.

Types and power of lights: Light source intensity is related to the rated power(e.g. in the corridor tested the lighting tubes are rated 35 W whilst theadjoining offices are fitted with 58W tubes). Manufacturers offer a widerange of light colours (e.g. Cool White) that might form a basis for dis-tinguishing electrical light sources based on their different spectra.

Radiant intensity variances: Ageing effects of fluorescent tubes usually go un-noticed by the human eye, although an intensity reduction of ∼20% canbe expected in 2 years in an office environment. Intensity decrease can beexacerbated by dirt on the tubes, so that a loss of ∼30% intensity in sixmonths can be found in a garage for example.3 The measurements made inSection 4 showed that the tubes tested had measurable (5%) differences inintensity. However, we do not anticipate that these differences would forma basis for identifying tubes that could be scaled up e.g. from a corridorto a building.

Light tube ballast identification: By modifying the fluorescent tube ballast,such lights have been used as a data broadcast channel. A number ofscenarios can be found at [20]. Location tracking using fluorescent lightsources might be simplified by having ballasts that allow each light totransmit a unique signature. A similar signature could be provided by LEDlighting; this approach might provide wireless broadband (up to 1 GB/s)transmission capacity [21].

2.2 Issues Related to the Tracking Environment

Objects and wall reflection: Here we consider the influence of surroundingse.g. walls and objects. Their impact will depend on the distance to theobjects from the light source(s), their shape and their reflectivity acrossthe spectrum. Generally reflection contributes insignificantly to the overallindirect illumination [22]. Since this is a static property of a buildingsection, it could be accounted for in building maps.

Availability of mapping information: It is understood that location systemsthat do not rely on such information being provided at the outset aregenerally preferred. Should map information be available, this informa-tion supports the tracking system by providing location constraints, e.g.impossible to walk through rigid objects such as walls. In this work we

3 Personal communication of the authors with a large Swiss light installation com-pany.


assume the type and radiation intensity of the building lights is knownfor the system calibration.

Influence of natural light from windows: The tracking of artificial lights maybe influenced significantly when solar cells are exposed to natural daylightnear windows. However measurements showed that intensity of the naturallight decays exponentially with the distance from the window and thewindow itself may have a transmissivity of 50% or less [22]. Hence at adistance of a few meters (depending on building and window form factor),the observed natural light intensity is lower than the indoor indirect lightintensity4 as shown in Fig. 2.

0 2 4 6 8 1010

0

101

102

Distance from window [m]

Irra

dian

ce [

W/m

2 ]

Daylight sourced irradianceFluorescent tube sourced irradiance

Fluorescent tubes

Fig. 2. Comparison of indoor irradiance source measurements, based on Fig. 3.28in [22].

In all light tracking experiments direct natural light was avoided by simplykeeping at least 3meters away from windows. However the windows couldbecome a data source for tracking. In a case with daylight and artificiallight from fluorescent sources only, e.g. an office building, natural lightmight be distinguished as it does not have the characteristic frequencyvariation (e.g. 100 Hz in Europe).

4 This assumes that direct light is partially transmitted or filtered by the window.


2.3 User Related Influences

Body constitution: Each user has a different physical constitution and bodyheight which affects the distance to overhead lights. These differences leadto small offsets in the solar cell voltage as indicated in Section 3.

Movement: As the user moves, the position of the solar cells with respect tothe body centre of mass will change e.g. due to user gait. Initial investi-gations of gait influence are made in Section 4.

Shadowing effects: The head and/or the body can block light paths, reducingthe intensity measured at the solar cell. Hence we propose the shouldersas the primary place for mounting solar cells as these positions tend to beleast obscured.

Clothing influence: For the case of solar cells on the shoulder, the orienta-tion of solar cells with respect to the horizontal plane can be affected bylooseness of clothing. For the current work a rigid type solar cells wasused.

3 Light Intensity Analysis and Simulation

3.1 Radiant Intensity Modelling.

In this first theoretical approach, we rely on radiant intensity informationextracted from artificial light sources only. More precisely, in this analysiswe consider a hallway scenario equipped with regular fluorescent light tubesinstalled in the ceiling at 2.5 m distance from the floor.

Many emitters can be modelled as a point source at sufficient distance.In this section we apply a point source to depict the principle propertiesof light sources. However, since we aim to investigate the radiant energy ofa light tube at different distances the model of a single point source is notsufficient. Therefore the fluorescent light tube was approximated as a boundedconcatenation of point sources (Section 3.3).

Light Emitter Model:

The source of radiant energy (emitter) creates a field of radiant flux. The totalreceived flux per area is called radiant intensity (in W/m2).

As the distance between light source and photovoltaic solar cell (receiver)changes, so does the radiant intensity received at the cell (radiant intensity).Radiant intensity IRPS at a distance r from a point source emitting radiantenergy with intensity I0 is related by the inverse square law [22]:

IRPS =ct ∗ I0

r2with ct = const. and I0 = const. (1)


Fig. 3. Schematic for the light emitter model

For positioning in three-dimensional space the coordinate system x, y, zshown in Fig. 3 will be used, with its origin at the centre of the light source0(x0, y0, z0).

In the particular case where a receiver is positioned on a plane withzS = const. near to the light source, the radiant intensity is at maximum ifxS = x0 and yS = y0. This can be described as the total planar radiantintensity ITPS . In the general case, for the radiant intensity at the solar cellsensor ISPS an arbitrary angle ϕ must be considered, with −90 ≤ ϕ ≤ 90

between the point source emitter and the receiving sensor, related by thecosine law: ISPS(ϕ) = IRPS ∗ cos(ϕ).

The distance r can be decomposed in the coordinate system by the threecoordinates positioning the solar cell S(xS , yS , zS) (see Fig. 3) depending onthe position along the light tube (y-coordinate).

r(y) =√

xS2 + (yS − y)2 + zS

2 (2)

In this representation ϕ is absorbed into r(y). The radiant intensity from apoint source ISPS can be directly related to the coordinate system:

ISPS(y) =c ∗ I0 ∗ zS

(xS2 + (yS − y)2 + zS)3/2

(3)

This relation can be used to simulate the radiant intensity received by ahorizontal solar cell moving in any direction under light sources at arbitraryheights. The total radiant intensity IS for a fluorescent light tube is found byapproximating the light as a chain of point sources in y-direction:

IS =

yL∫−yL

ISPS dy =

yL∫−yL

c ∗ I0 ∗ zS

(xS2 + (yS − y)2 + zS)3/2

dy (4)

For simulations the distance to the light source, e.g. z-component zS ofS can be assumed to be constant. From the perspective in x-direction, the


model assumes the light tube as a point source. Hence, xS is constant. Thetotal length of the light tube is denoted by l. Hence, the integration limits aredescribed by yL = l/2.

3.2 Detection of Light Emitters

The change in radiant intensity when varying the distance of a photovoltaicsolar cell to the light emitter can be monitored by current or voltage variation.Whilst current is directly proportional to radiant intensity, voltage varies withthe log of intensity: V ∝ ln(IS). For the simulation, the estimates c = 1 andI0 = 5W/m2 were used. These values are fitted with real measurements toreflect size and type of the solar cells in the model.

1 2 3 4 5 6 7 8 91

1.5

2

2.5

3

3.5

x in [m]

Log

radi

ant i

nten

sity

zs=0.6 mzs=0.8 mzs=1.0 m

Fig. 4. Simulation of a movement along x under three point sources in 2.5m spacingat different ceiling distances zS of S (yS = y0).

Fig. 4 shows the expected waveform for a straight trajectory equidistantto the walls down a corridor. The light tubes are oriented at right angles tothe trajectory and regularly distributed. Since for this simulation example they-component of the movement is ∆y = 0, the light tubes are approximatedas point sources. Assuming a typical office building height of about 2.5m, thedistance zS from an adult shoulder to the ceiling mounted light sources willbe about one meter or less. The regular distance dL between the light tubesis greater than 2.5 meters. Using the light emitter model a lower bound onthe amplitude fluctuation can be derived: for the case of dL = 2.5 m andzS = 0.8 m there is a difference of 20% in the radiant intensity from theminimum to maximum value. This difference is deemed sufficient to correctlydetect when the solar module is beneath a tube.


Environment Effects and Emitter Model Limitations:

The light emitter model does not consider indirect light, such as reflectionsfrom walls and cupboards. As indicated in Section 2 indirect light generallyhas an order of magnitude less intensity than direct light, so this componentof radiant energy has not been included in the light emitter model.

Occluded direct light can create distinct shadows hindering radiant energyreaching a sensor. For the intended application using overhead light sources,the possible obstruction area is limited to objects in direct line of sight withthe light source, e.g. the box frame supporting the fluorescent light tube orthe user’s head. As this model does not cover human aspects in detail, thehead is not considered.

3.3 Light Emitter Based Distance Model

In a second modelling step, a 3-dimensional environment for configurable lightdistributions was built for the majority of the corridors of our offices. Withthis approach it is possible to simulate various building scenarios. It is usedhere exclusively for a section of the corridor scenario.

For the following analysis of a hallway in an existing office building, theradiant energy distribution of fluorescent light tubes has been simulated asshown in Fig. 5. The simulation covers one long corridor leading to officesand a connecting passage. The simulated portion has 3 identical light tubes,equidistant with both walls, oriented in a perpendicular direction to the maincorridor access. The distance between the lights is dL = 3.7 m. There are noother significant sources of light in the scenario.

4 Light Source Tracking and Activity Estimation

4.1 Experiments

The initial goal of the experiments was to verify the waveforms calculated inthe simulation. At the same time, measurement data were acquired for creat-ing and validating an estimation model. This section details the measurementsystem used and the data sets acquired. Experiments were carried out withinthe corridor in which radiant energy had been simulated.

Sensing System:

A measurement system based on a trolley was built that allows acquisition ofthe voltage from solar cells. The same set up was used for all experiments inthis section. The solar module was positioned in a horizontal plane on top ofthe trolley. Furthermore, a relatively constant distance zS between cell and


Fig. 5. Example of a simulated radiant intensity distribution as 3D plot for thescenario

fluorescent light tubes was maintained that varied by a few millimetres dueto ground roughness under the wheels of the trolley for example.

The photovoltaic solar module used for the experiments is an amorphoussilicon thin film deposited on glass.5 The voltage of the solar cells was acquiredat 1 kHz and 12 bit resolution using a standard data acquisition system. Forconvenience of signal acquisition, solar cell voltage across a 10 kΩ resistancewas tracked in all experiments.

Verification Method:

To associate the acquired waveforms with actual distance down the corridor,two approaches were used: firstly we made measurements between manuallylocated tags. This technique whilst relatively precise was limited in resolutionsince data could only be collected manually every ∼50 cm.

Therefore a second distance measurement method was used: the frontwheels of the trolley were replaced with bicycle wheels fitted with a standardbicycle dynamo. A peak detection of the acquired voltage waveform from thedynamo generator was used to determine the system speed. The accuracy ofthe wheel measured distance compared to the distance determined by the firstmethod was always over 98%. This accuracy was deemed sufficient and thesecond method was used to acquire ground truth for subsequent experiments.

5 Manufacturer: RWE SCHOTT Solar, model: ASI 3 Oi 04/057/050


Description of Experiments:

The experiments were performed pushing the trolley at constant walkingspeed (0.55 ms−1) in an office corridor. A straight trajectory was taken belowthe middle of the light tubes. These tubes were integrated into the ceiling andoriented perpendicular to the walls of the corridor. The light diffusers weremade of grooved (principle of Fresnel) colourless perspex. The distance be-tween the horizontal solar cell and the fluorescent light tubes was zS = 73 cm,the distance between the light tubes was dL = 3.7 m.

The solar module voltage across the resistance of 10 kΩ was collected overtrajectories of 2m to 10 m in length. Each time the solar module passed undera tube, a waveform peak was measured. A total of 14 such peaks were recorded.

Additional experiments were conducted to analyse if more detailed humanmotion than horizontal translation could be detected using solar cells. Theshoulder was selected for placing the solar cell as we believe that this bodyposition would be optimal with respect to maximising energy collected bythe solar cell as well as ensuring optimal recognition of overhead lighting. Wesought to quantify the effect of human walking motion as this would be aperturbation of a garment based distance measuring system.

A healthy male subject (69 kg) with body height 1.82m was selected whowore a close fitting T-shirt with Velcro sewn on the shoulders. The solar mod-ule was fixed onto the left shoulder and an acceleration sensor6 was attachedto the back at waist level (centre of mass). The subject walked on a treadmillthat was running at a fixed speed. The subject’s speed was kept constant bymaintaining a relatively fixed position on the treadmill. The solar cell wasdirectly beneath a fluorescent tube, similar to those installed in the corridorexperiments. The subjects position was also equidistant from the ends of thefluorescent tube. Once the subject was comfortable with the speed, the so-lar module and acceleration sensors were synchronised and monitored for 1minute. Data was collected at four speeds (0.28, 0.83, 1.39 and 2.78 m/s) ofthe treadmill.

4.2 Results

Light Emitter Model Validation:

To validate the light emitter model we compared the simulation result withthe measured waveforms of the hallway scenario. The light emitter modelvaried over the range of 0.4 to 3.3 V whilst the average of the measured valueswas in the range 0.1 to 2.7 V. The simulation error for ten peak waveformswas less than 0.35 V in 80% and 0.4 V in 90% of all measurements. Part ofthe error can be attributed to the emitter model being for a bare fluorescentlight tube rather than the measured data which was for an installed light tubeincluding a reflective backing and a front diffuser.6 Manufacturer: XSens Technologies B.V., model: MT9B


Displacement Estimation from the Solar Cell:

Distance information can be extracted from the amplitude of the waveform byusing a mapping between voltage and known distance from training data. Tobuild the model, training data sets were segmented into single peak waveformsand scaled to the known light to light distance of dL = 3.7 m. This empiricalmodel was used in a simple mapping function which relates observed voltagesfrom the solar cell and the average displacement under the light source.

To improve the displacement estimation for measurement outliers duringtesting that were not covered by the trained model, the last available speedinformation obtained with the same model was used. The average positionestimation error (APEE) obtained with this method is less than 21 cm with aconfidence of 80% over a distance of 7.4m. The absolute position estimationerror is depicted in Fig. 6.

0.05 0.1 0.15 0.2 0.250

10

20

30

40

50

60

70

80

90

100

APEE [m]

Perc

enta

ge o

f Cas

es

Fig. 6. Absolute position estimation error (APEE) of the empirical model

Human Activity Recognition:

The vertical displacement of the human shoulder during locomotion is lessthan 10 cm. A point on the human shoulder makes a movement while walkingthat resembles a U shape when viewed from the front. This is the case forspeeds from 1.6–2.3 m/s [23]. This displacement can be modelled by an overlayof harmonic sinusoidal functions.

A direct relationship in frequency and phase between the acceleration inthe z-axis (up and down movement) and the solar module voltage was ob-served. We deduce that the cyclical trace at the solar cell is related to thestriding motion of the subject [24]. Good correlations are achieved at tread-mill speeds of 1.39 m/s and below. At higher speeds (e.g. when the subject is


running), the estimation method incurred large errors and outliers. Further-more the walking waveforms showed some outliers for higher speeds that couldbe explained by abnormalities in the rhythmic movements, e.g. small jumpsor bends of the torso. The peak to peak voltage variation induced by walkingis low, e.g. 0.03 V at 1.39m/s, suggesting satisfactory position tracking whilewalking can be expected as outliers are automatically ignored by the empiricalmodel described above. Since walking introduced a low cell voltage variation,also a low influence of walking on the position estimation by light tubes canbe expected since outliers are automatically ignored by the empirical modeldescribed above.

Whilst a sinusoidal model resembles the solar cell measured trace, it is nota conclusive match, since the model does not include rotational (forward andback motion and swing) of the shoulder.

5 Extension of LuxTrace with Absolute PositioningInformation

We present our initial results on the extension of the LuxTrace system withabsolute positioning systems. The motivation behind this investigation weretwofold: 1) we evaluate an automatic online learning and integration strat-egy with absolute positioning systems and 2) we intended to gather moreevaluation data.

5.1 System Combination Approach

The obvious motivation for combining positioning systems is reliability. A sin-gle system may fail or be locally unavailable for various reasons, see the issuesrelated to the LuxTrace approach in Section 2. Another important feature isadaptability. The combination of well-selected position sensing methods canfacilitate the system’s adaptability to the current location. For LuxTrace thiscan be helpful, e.g. in interconnected buildings with different lighting schemes.

To evaluate the adaptability concept we use an online procedure for learn-ing LuxTrace models with the help of additional absolute positioning informa-tion. The absolute positioning reference used in this work are carpet-like RFIDfoils distributed on the floor. These foils provide positioning information bythe unique identification integrated in the tags that can be read when passingover them. The combined system consists of three parts: 1) the LuxTrace solarcell, 2) the RFID reading system and 3) the model learning and displacementestimation module. Fig. 7 depicts the system architecture of the combined ap-proach. The solar cell and RFID reading parts acquire sensed data; the modellearning and estimation module is used for training and displacement esti-mation. The training of a LuxTrace displacement model is initiated wheneversufficient RFID readings are available. In this work we used the observation ofone light period for the training, hence the RFID service must be available for


individual lights only. The trained model is used immediately for the estima-tion. With this approach a sparse RFID distribution is possible, supportingthe estimation of the solar cell at critical locations.

RFID reader moduleSolar cell module

model learning anddisplacement estimation

Fig. 7. Modules and signal flow in combined system.

We selected RFID tags integrated in carpet-like foils as complementarytechnology for a number of reasons. Firstly, RFID technology offers an abso-lute reference positioning service with low-maintenance. Secondly, RFID is aninexpensive technology that is widely used. Thirdly, we could use the mem-ory capacity of the RFID tags for providing additional location-dependentinformation in-situ.

5.2 Experiments

Sensing System:

The sensor hardware and the A/D converter experiments are the same asthose used during the original LuxTrace experiments.

We have used an existing RFID positioning system developed at ETHZurich as the absolute positioning service [17]. It consists of two major com-ponents: 1) a trolley equipped with an off-the-shelf RFID tag reader and anRFID antenna, that is mounted at the bottom of the vehicle and 2) a carpet-like RFID tag infrastructure [25] that provides a dense distribution of RFIDtags across certain areas of the floor space. The basic operating principle ofthe RFID positioning service is as follows: tag position coordinates stored onindividual RFID tags are read that are within antenna range. The current po-sition is computed as the arithmetical mean of the individual position readingsobtained.

The LuxTrace equipment was added to the RFID positioning trolley. Thecomplete system is shown in Fig 8. The RFID antenna was attached underthe base, at 10 cm above the floor space. At this distance, the approximatelysquare operating area of the RFID antenna was about 50 × 50 cm. RFIDtagged foils with an average tag density of 39 tags/m2 were deployed.


Black Box with Solar Module

Notebook executing Combined Systemand RFID Positioning applications

Trolley

A/D Convertor

12V-Battery for RFID Reader

RFID Reader

RFID Template

Notebook acquiring Solar Cell Voltages

Fluorescent Light Tubes

RFID Antenna (underside)

Fig. 8. Measurement trolley and prepared test track.

For our practical experiments we chose an office corridor of similar typeas the one used for the LuxTrace series of experiments, with light tubes ofidentical type, arrangement, and light emitting characteristics. We installedfour RFID-tagged foil templates along the length of the corridor. The layoutof our experimental setup is shown in Fig. 9. The solar cell was mounted onthe trolley at 93 cm from the ceiling of the office corridor.

(0,0) (123,0) (370,0) (493,0)

(123,128) (246,128) (493,128) (616,128)

RFID-Template 1 124

370 370

Trajectory, 10 m

x

y

RFID-Template 2

RFID-Template 3

RFID-Template 4

Fig. 9. Experimental setup: position of light tubes and RFID tagged foil templates(all coordinates and distances in cm). The test track starting at position (0,60) islocated at the start of the red arrow (trajectory 10m).

Description of Experiments:

We limited the observation range during the model learning procedure to asingle low-high-low cycle of light intensities, as it is desirable to minimise thelength of the corridor needed for the training of new models.

The placement of the RFID templates was chosen in a way that allowed usto obtain RFID position information in the critical sections halfway between


adjacent light tubes, that can be used to support detection of low points inthe measured light intensity.

All practical experiments were performed by pushing the measuring trolleyat a constant walking speed of ∼0.34m/s along the test track.

In order to assess the positioning errors of the combined system under re-alistic conditions, each trained model would ideally be validated by comparingits position estimates to the actual trajectory along the test track a numberof times, exposing it to different light intensity trace for each pass. For practi-cal reasons, we experimentally measured and recorded multiple light intensitydata streams along the same trajectory. We then replayed these recorded ex-perimental data streams in a simulation mode and compared them with thedifferent recorded reference positioning data streams. We used this methodto simulate further experiments and thus obtain additional data by cross-combining of data streams from different experiments.

5.3 Results

Displacement Estimation of the Combined System:

We compared our results for the combined system with those previously re-ported in Section 4.

0 0.1 0.2 0.3 0.4 0.5 0.6 0.70

10

20

30

40

50

60

70

80

90

100

APEE [m]

Perc

enta

ge o

f Cas

es

LuxTracecombined system

Fig. 10. Cumulative absolute positioning estimation errors of both investigatedapproaches: original LuxTrace model and combined system results.

Figure 10 depicts the measured cumulative absolute positioning estimationerrors (APEE) for the combined system and the original LuxTrace results forreference purposes.

Based on this plot, we can make the following observations: for the quan-tiles above 90%, the combined system has a significantly lower performance in


comparison to the original LuxTrace result. However, for all quantiles ≤90%,the maximum error of the combined system never exceeds the maximum errorof the LuxTrace reference system by more than 8 cm. For the 80% and 90%quantiles, the accuracy of the position estimates remain below 28 cm and 33 cmrespectively for the combined system, and below 21 cm and 25 cm respectivelyfor the original model. We consider these results an indication of the robust-ness of the online model learning approach, which merely relies on the analysisof a single light period in an arbitrary corridor. In contrast, the original Lux-Trace model was trained with the data from 10 light period measurements fora particular corridor.

6 Discussion

We assess our results with respect to the location technology assessment tax-onomy proposed by Hightower in the Location Stack project [26] that consid-ers scalability, cost, recognition and limitations.

Scalability: A scalability factor related to cost is that as the solar cell de-vices and RFID tags are relatively cheap, it can be anticipated that thenumber of these devices, both solar cells on the body and RFID tags inthe environment, should not be a limiting factor. For the solar cell basedLuxTrace a wide application range in offices and manufacturing facilitiesis expected.The combined system improves on the scalability of LuxTrace, as fullcoverage with both the solar cell and RFID systems is not required. Also,as RFID tags are fixed, they help to reduce the risk of drift from the solarcells.

Cost: For the LuxTrace system consisting of a solar cell and acquisition unit, alow cost implementation can be achieved. The global cost of the combinedsystem includes incremental hardware, installation and maintenance. Fora wearable device, we would anticipate that one or more antenna builtinto the shoe(s) would be used. This receiving device would be the highestpower (Watt) incremental sensing device, and overall design would seekto minimise its duty cycle by switching when possible to the solar cellsystem (Milliwatt) only. The corollary of such reduced power consumptionis reduced mass, volume and cost for the battery; such components oftenhave the highest values of the latter criteria in mobile systems.A scenario minimising the installation cost of the RFID tags could beachieved if they were delivered as part of the carpet in a new building.Such tags could also be retrofitted under the existing carpet by using foilsas used in the experiments of this work. A further way to minimise costwould be to install the tags only at intersections (e.g. corridor-corridor orcorridor-room).Further hardware costs are avoided as indoor lights and on-body comput-ers (e.g. mobile phone) are generally available; incremental maintenance


costs of the system would be minimal assuming that lighting infrastruc-ture is not significantly changed and a bulb replacement service(s) exist.

Recognition: We have shown that recognition has not suffered despite themore practical scenario of the online model training and position estima-tion. Furthermore, combining the solar cell and RFID system improvesreliability by supporting calibration of one system against the other. Inthis way our work presents a simple but robust feature fusion approachof the two systems. We manage the navigation system energy consump-tion and accuracy as follows: for maximum accuracy we can use bothsystems whilst for energy saving or for wearable computing scenarios, weselectively use only the lowest power (solar cell based estimator) system.

Limitations: The reported experiments focused exclusively on indoor appli-cations in the absence of natural light. While this is sufficient for manyroom and corridor scenarios further work is necessary to determine therestrictions incurred from natural light sources.

7 Further Work

Our hardware is not adapted to wearable use and could be improved in anumber of ways including, but not limited to, incorporating the RFID an-tenna into a pair of shoes, using flexible solar cells rather than glass depositedones and experimenting with the use of multiple solar cells on the same user.Further experiments with humans would allow us to investigate more realisticscenarios, trajectories and location tracking durations.

In order to improve the wearability of the concept, the use of flexiblesolar cells (e.g. from [27]) would better match garment specification thanthose used in the experiments presented. Our results support the feasibility offuture sensor nodes that rely on the same device for both sensing and energyharvesting.

Our models could also be improved with probabilistic algorithms, e.g.Kalman or particle filters. For the shoulder movement that we have inves-tigated, a model based on rotations of a point on the shoulder in 3D spacecould be considered. Success in modelling such sources of context could alsofurther support recognition accuracy.

It would be useful if LuxTrace could provide orientation around the verticalaxis of horizontal solar cell. This might be inferred if the solar cells and lightsources were equipped with polarising filters.

The combination of a low-power light sensing device (power requirementsfor A/D converter and voltage transformer: ∼7 Milliwatt) with a relativelyhigh power RFID system (power requirements: ∼1 Watt) serves as an exam-ple of information fusion trade-off between coverage, accuracy and energy ef-ficiency. The system could therefore adapt to maximum accuracy mode usingthe high-power RFID positioning system; an energy saving (wearable) mode


would selectively use only the low power system. For maximum availabilityand coverage, both systems would be used in parallel.

8 Conclusion

We presented a concept to support location tracking with solar cells for indoorwearable applications. In our approach radiant energy from indoor buildingillumination is monitored by solar cells to derive displacement and contextinformation.

After theoretical evaluation we presented experimental results indicat-ing that a displacement estimation accuracy of 21 cm (80% quantile) can beachieved on straight trajectories. Additionally we evaluated the influence ofhuman walking on the solar cell received irradiance.

In the second step we briefly investigated the collaboration of our systemwith an absolute positioning system. We developed a combined system con-sisting of a displacement estimator based on light sensing and informationcollected from carpet-like RFID tags distributed in the environment. Thedisplacement estimator learns new lighting schemes online if RFID data isavailable. We found that the accuracy of the combined systems was similar(28 cm for 80% of all measurements).

We conclude that for the controlled conditions in which we experimented,LuxTrace (i.e. solar cells alone) are generally sufficient. An example appli-cation would be a linear displacement measurement on a track. However, forlocation tracking applications with less prior knowledge (e.g. spacing of lights,starting position, less predictable trajectories) the combined system may provemore accurate.

References

1. Getting, I.: The Global Positioning System. IEEE Spectrum 30 12 (1993) 36–472. Bretz, E.A.: Precision navigation in European skies. IEEE Spectrum 40 (2003)

163. European Commission: Galileo project webpage. http://www.europa.eu.int/

comm/dgs/energy_transport/galileo/index_en.htm (2005)4. European Commission: The Galilei Project: GALILEO Design Consolidation.

http://europa.eu.int/comm/dgs/energy_transport/galileo/doc/ (2003)5. Ni, L.M., Liu, Y., Lau, Y.C., Patil, A.P.: LANDMARC: indoor location sensing

using active RFID. Wirel. Netw. 10 (2004) 701–7106. Vossiek, M., Wiebking, L., Gulden, P., Weighardt, J., Hoffmann, C.: Wireless

local positioning - concepts, solutions, applications. In: Radio and WirelessConference. (2003) RAWCON ’03.

7. Agiwal, A., Khandpur, P., Saran, H.: LOCATOR: location estimation systemFor wireless LANs. In: WMASH ’04: Proceedings of the 2nd ACM internationalworkshop on Wireless mobile applications and services on WLAN hotspots, NewYork, NY, USA, ACM Press (2004) 102–109


8. Andy, W., Alan, J., Andy, H.: A New Location Technique for the Active Office.IEEE Personal Communications 4 (1997) 42–47

9. Vildjiounaite, E., Malm, E., Kaartinen, J., Alahuhta, P.: Location estimationindoors by means of small computing power devices, accelerometers, magneticsensors, and map knowledge. In: Pervasive ’02: Proceedings of the 1st Interna-tional Conference on Pervasive Computing, Springer (2002) 211–224

10. Lee, S., Mase, K.: A personal indoor navigation system using wearable sensors.In: Proceedings of the 2nd International Symposium of Mixed Reality (ISMR01),Yokohama (2001)

11. Randell, C., Muller, H.: Low cost indoor positioning system. In: Abowd, G.D.,Springer (2001) 42–48

12. HeyWow Project: HeyWow project webpage (2005) http://www.heywow.com.13. Bohn, J., Vogt, H.: Robust Probabilistic Positioning Based on High-Level

Sensor-Fusion and Map Knowledge. Technical Report 421, Institute for Per-vasive Computing, Dept. of Computer Science, ETH Zurich, Switzerland (2003)

14. Angermann, M., Kammann, J., Robertson, P., Steingass, A., Strang, T.: Soft-ware representation for heterogeneous location data sources within a probabilis-tic framework. In: International Symposium on Location Based Services forCellular Users (Locellus). (2001) 107–118

15. Kargl, F., Bernauer, A.: The COMPASS Location System. In: LoCA. (2005)105–112

16. Hightower, J., Brumitt, B., Borriello, G.: The Location Stack: A Layered Modelfor Location in Ubiquitous Computing. In: Proceedings of the 4th IEEE Work-shop on Mobile Computing Systems & Applications (WMCSA 2002), Callicoon,NY, IEEE Computer Society Press (2002) 22–28

17. Bohn, J.: Prototypical Implementation of Location-Aware Services based onSuper-Distributed RFID Tags. In: Proceedings of the 19th International Con-ference on Architecture of Computing Systems (ARCS’06), Frankfurt am Main,Germany. Number 3894 in LNCS, Springer (2006) 69–83

18. Amft, O., Lauffer, M., Ossevoort, S., Macaluso, F., Lukowicz, P., Troster, G.:Design of the QBIC wearable computing platform. In: Proceedings of the 15thIEEE International Conference on Application-specific Systems, Architecturesand Processors. (2004)

19. Fox, D.: Markov Localization: A Probabilistic Framework for Mobile RobotLocalization and Navigation. PhD thesis, Institute of Computer Science, TUDreseden, Germany (1998)

20. Lights, T.: Talking Lights, Boston MA USA. http://www.talking-lights.

com/how.htm (2005)21. Kavehrad, M., Amirshahi, P.: Hybrid MV-LV Power Lines and White Light

Emitting Diodes for Triple-Play Broadband Access Communications. In: IECcomprehensive report on Achieving the Triple Play: Technologies and BusinessModels for Success, ISBN: 978-1-931695-37-4. IEC (2005) 167–178

22. Randall, J.: Designing Indoor Solar Products. John Wiley and Sons Ltd, ISBN0-470-01661-2 (2005)

23. Cappozzo, A.: Analysis of the linear displacement of the head and trunk duringwalking at different speeds. Journal on Biomechanics 14 (1981) 411–425

24. Amft, O., Randall, J., Troster, G.: Towards LuxTrace: Using solar cells tosupport human position tracking. In: Proceedings of 2nd International Forumon Applied Wearable Computing, Zurich, Switzerland. (2005)


25. Bohn, J., Mattern, F.: Super-Distributed RFID Tag Infrastructures. In: Pro-ceedings of the 2nd European Symposium on Ambient Intelligence (EUSAI2004), Eindhoven, The Netherlands. Number 3295 in LNCS, Springer (2004)1–12

26. Hightower, J., Borriello, G.: A survey and taxonomy of location systems forubiquitous computing. IEEE Computer 34 (2001) 57–66

27. VHF Technologies: Yverdon-les-Bains Switzerland. http://www.flexcell.ch

(2005)

LuxTrace - Indoor Positioning Using Building Illumination · perposition for energy eﬃciency and...

Documents

Transcript of LuxTrace - Indoor Positioning Using Building Illumination · perposition for energy eﬃciency and...